Summary
Managed tools are compiled into per-agent tools.json manifests, but OpenClaw agents behind cllama passthrough still do not actually receive or see those tools at runtime.
In Tiverton, claw up -d succeeds, .claw-runtime/context/<agent>/tools.json is populated, and the live pod is healthy. But agents still report that they do not have MCP trading tools and fall back to wrapper scripts.
Expected
Given a service descriptor with version: 2 tools[] and a pod service subscribed via x-claw.tools, the agent should be able to call those managed tools through the normal model tool-calling path.
At minimum, one of these needs to be true:
cllama loads tools.json from /claw/context/<agent>/tools.json and injects provider-native tools[] into chat requests before forwarding upstream, or- the runner config (
openclaw.json here) is materialized with the managed tools locally so the runner sends them itself.
Actual
tools.json is compiled correctly for Weston:
/home/clawdbot/tiverton-house/.claw-runtime/context/weston/tools.json- contains tools like
trading-api.propose_trade, trading-api.get_market_context, etc.
- Weston's generated OpenClaw config contains only the built-in web search tool:
/home/clawdbot/tiverton-house/.claw-runtime/weston/config/openclaw.jsontools.web.search.provider = "perplexity"- no managed trading tools appear there
- The running
cllama container mounts the context root at /claw/context
- The live
cllama logs show feed injection activity (feed_fetch) but no evidence of tool loading/injection/execution for Weston
- Weston explicitly reported in Discord that he has no MCP trading tools in his tool list and is falling back to mounted scripts
Evidence from current source
The current OpenAI handler in cllama appears to inject feeds and current time only:
cllama/internal/proxy/handler.go
fetchFeeds(...)feeds.InjectOpenAI(payload, feedBlock)feeds.InjectOpenAI(payload, currentTimeLine(...))
I do not see any adjacent logic that loads tools.json, injects tools[], intercepts tool calls, or executes managed HTTP tools.
This also seems to conflict with the current Clawdapus changelog, which says:
- compiled
tools.json is written into cllama context
- managed tool mediation in cllama is supported
Repro
- Configure a service with a v2
claw.describe containing tools[]
- Subscribe an OpenClaw agent via
x-claw.tools
- Run
claw up -d
- Confirm
.claw-runtime/context/<agent>/tools.json exists and contains the expected tools
- Ask the live OpenClaw agent whether it has those tools available
Observed environment
- host
claw version: 0.5.1
cllama image: ghcr.io/mostlydev/cllama:latest- OpenClaw image:
ghcr.io/openclaw/openclaw:2026.3.24
- live pod: Tiverton House
Impact
This blocks the main user-visible goal of the managed-tools rollout for OpenClaw pods: traders are told to prefer tools, the manifests are generated, but the tools are not actually reachable in the agent runtime.
Related
This is separate from #112, which is about descriptor discovery for build-only services without image:. That issue was worked around and the pod is otherwise live.
Summary
Managed tools are compiled into per-agent
tools.jsonmanifests, but OpenClaw agents behindcllamapassthrough still do not actually receive or see those tools at runtime.In Tiverton,
claw up -dsucceeds,.claw-runtime/context/<agent>/tools.jsonis populated, and the live pod is healthy. But agents still report that they do not have MCP trading tools and fall back to wrapper scripts.Expected
Given a service descriptor with
version: 2tools[]and a pod service subscribed viax-claw.tools, the agent should be able to call those managed tools through the normal model tool-calling path.At minimum, one of these needs to be true:
cllamaloadstools.jsonfrom/claw/context/<agent>/tools.jsonand injects provider-nativetools[]into chat requests before forwarding upstream, oropenclaw.jsonhere) is materialized with the managed tools locally so the runner sends them itself.Actual
tools.jsonis compiled correctly for Weston:/home/clawdbot/tiverton-house/.claw-runtime/context/weston/tools.jsontrading-api.propose_trade,trading-api.get_market_context, etc./home/clawdbot/tiverton-house/.claw-runtime/weston/config/openclaw.jsontools.web.search.provider = "perplexity"cllamacontainer mounts the context root at/claw/contextcllamalogs show feed injection activity (feed_fetch) but no evidence of tool loading/injection/execution for WestonEvidence from current source
The current OpenAI handler in
cllamaappears to inject feeds and current time only:cllama/internal/proxy/handler.gofetchFeeds(...)feeds.InjectOpenAI(payload, feedBlock)feeds.InjectOpenAI(payload, currentTimeLine(...))I do not see any adjacent logic that loads
tools.json, injectstools[], intercepts tool calls, or executes managed HTTP tools.This also seems to conflict with the current Clawdapus changelog, which says:
tools.jsonis written into cllama contextRepro
claw.describecontainingtools[]x-claw.toolsclaw up -d.claw-runtime/context/<agent>/tools.jsonexists and contains the expected toolsObserved environment
claw version: 0.5.1cllamaimage:ghcr.io/mostlydev/cllama:latestghcr.io/openclaw/openclaw:2026.3.24Impact
This blocks the main user-visible goal of the managed-tools rollout for OpenClaw pods: traders are told to prefer tools, the manifests are generated, but the tools are not actually reachable in the agent runtime.
Related
This is separate from #112, which is about descriptor discovery for build-only services without
image:. That issue was worked around and the pod is otherwise live.